@logicpearl/try

Kick the tires on LogicPearl in one command. No setup, no API keys, no config — the package ships with a real compiled Wasm artifact and five real captured Claude Sonnet 4.5 responses to the same input.

npx @logicpearl/try

You get a deterministic refund verdict in under a millisecond. Same input → same bitmask → every run.

The demo

# The default — evaluate the shipped refund-eligibility artifact
# against a day-30 changed-mind boundary case.
npx @logicpearl/try

# Compare the artifact against 5 real captured Claude Sonnet 4.5 runs.
# (Ships with the captures — no API key needed.)
npx @logicpearl/try --explain

# Reproduce those captures live against your own key.
ANTHROPIC_API_KEY=sk-ant-... npx @logicpearl/try --explain --live

# Run the 3-stage pipeline on a plain-English customer message.
npx @logicpearl/try --from-text "I bought this a month ago and want a refund"

# Run that pipeline 5 times to see which stages vary and which stay identical.
npx @logicpearl/try --from-text "..." --prove-it

Flags

| Flag | What it does | |---|---| | --facts '{...}' | Evaluate your own feature vector | | --facts-file <path> | Load the facts from a JSON file | | --artifact <path-or-url> | Use a different compiled artifact | | --describe | Print the feature + rule schema only | | --explain | Side-by-side vs 5 pre-captured LLM runs | | --live | Call the LLM now (needs ANTHROPIC_API_KEY or OPENAI_API_KEY) | | --from-text "<prose>" | 3-stage pipeline: LLM extract → LogicPearl decide → LLM verbalize | | --prove-it | Run the pipeline 5× and report stage variance | | --json | Machine-readable output |

What's in the package

• A compiled refund-eligibility artifact (8 KB Wasm, 5 learned rules, 8 features including a string-categorical)
• captures/claude-sonnet-5-runs.json — five real API responses, timestamps, prompt hash
• captures/policy.md — the policy text the LLM was given
• captures/prompt.txt — the exact prompt template

Everything in captures/ is reproducible: pass CAPTURE_MODEL=claude-sonnet-4-5 ANTHROPIC_API_KEY=... node scripts/capture_llm_runs.mjs (in the repo) to regenerate them.

What's next

• @logicpearl/mcp — one-command MCP server install for Claude Desktop + Cursor
• @logicpearl/browser — loadArtifact() + evaluate() in the browser or worker
• Docs | Source

License

MIT